We consider the problem of using nearest neighbor methods to provide a conditional probability estimate, P(y|a), when the number of labels y is large and the labels share some underlying structure. We propose a method for learning error-correcting output codes (ECOCs) to model the similarity between labels within a nearest neighbor framework. The learned ECOCs and nearest neighbor information are used to provide conditional probability estimates. We apply these estimates to the problem of acoustic modeling for speech recognition. We demonstrate an absolute reduction in word error rate (WER) of 0.9% (a 2.5% relative reduction in WER) on a lecture recognition task over a state-of-the-art baseline GMM model.

Discriminative language models using n-gram features have been shown to be effective in reducing speech recognition word error rates. In this paper we describe a method for incorporating discourse-level triggers into a discriminative language model. Triggers are features identifying re-occurrence of words within a conversation. We introduce triggers that are specific to particular unigrams and bigrams, as well as "back off" trigger features that allow generalizations to be made across different unigrams. We train our model using a new loss-sensitive variant of the perceptron algorithm that makes effective use of information from multiple hypotheses in an n-best list. We train and test on the switchboard data set and show a 0.5 absolute reduction in WER over a baseline discriminative model which uses n-gram features alone, and a 1.5 absolute reduction in WER over the baseline recognizer.

Previous work has considered methods for learning projections of high-dimensional acoustic representations to lower dimensional spaces. In this paper we apply the neighborhood components analysis (NCA) [2] method to acoustic modeling in a speech recognizer. NCA learns a projection of acoustic vectors that optimizes a criterion that is closely related to the classification accuracy of a nearest-neighbor classifier. We introduce regularization into this method, giving further improvements in performance. We describe experiments on a lecture transcription task, comparing projections learned using NCA and HLDA [1]. Regularized NCA gives a 0.7 % absolute reduction in WER over HLDA, which corresponds to a relative reduction of 1.9%. Index Terms: speech recognition, acoustic modeling, dimensionality reduction